A unifying causal framework for analyzing dataset shift-stable learning algorithms

Abstract Recent interest in the external validity of prediction models (i.e., problem different train and test distributions, known as dataset shift ) has produced many methods for finding predictive distributions that are invariant to shifts can be used new, unseen environments. However, these consider types have been developed under disparate frameworks, making it difficult theoretically analyze how solutions differ with respect stability accuracy. Taking a causal graphical view, we use flexible representation express various shifts. Given graph data generating process, show all correspond hierarchy operators, which disable edges responsible The provides common theoretical underpinning understanding when achieved, what ways stable differ. We establish conditions minimax optimal performance across environments, derive new algorithms find distributions. By using this perspective, empirically demonstrate there is tradeoff between average performance.